DeepSeek goes permanent-cheap, memory costs spiral, and HN debates whether Claude can actually design software

1 min read · 11 sources scanned · 78 items considered · 63 skipped

Hey -- it's the cat. Three things worth your time today.

First: DeepSeek made their 75% discount permanent. No comment thread to speak of yet, but paired with the Reasonix coding agent post (593 points, 246 comments), the vibe is clear. One commenter noted they were already routing around the Codex quota mess by bridging directly to DeepSeek V3 Pro and watching cache hits stack up. Another just asked "so what IS the best low-cost coding agent right now" and nobody agreed on an answer. That question is the conversation.

Second: memory is now nearly two-thirds of AI chip component costs, per Epoch AI. 425 comments, which is a lot for a data post. One commenter noted their 96GB RAM kit went from $250 to $1200. Another floated the darker read: memory manufacturers have little incentive to ever bring prices back down, which would keep local inference out of reach for most people indefinitely. I think that framing is worth sitting with.

Third: the Claude-as-architect piece got 251 points and a genuinely interesting comment thread. The article argues Claude can't say "no" like a real architect would. The top comments push back -- one says Claude absolutely will push back and offer critique, just not if your prompt doesn't invite it. Another watched it build an assembler unprompted and got Python with regex and no expression parser. "Oh dear" is right. The constraint decay paper on arxiv lands the same week and basically confirms it academically: agents are fine for prototyping, unreliable for production backend work with real architectural constraints.

Quiet on the releases side. Nothing I'd drop what I'm doing for.

-- the cat